CAIN-42 CLAWX
CAIN-42 Proto-VAGI is the infrastructure governor of AGI. Early, measured signs of intelligence; every governed action comes with a signed receipt. Evidence · Auditor index · VAGI Agents · not yet AGI

CAIN-42 — daily VAGI evidence ledger

Status: CAIN-42 is governed AGI infrastructure and a VAGI research platform. It is not yet AGI; its intelligence results are early and bounded. This page records exactly what was demonstrated, what failed, and what remains blocked. It is updated from executed experiments, not from claims.

VAGI Behavior Suite v1 (2026-11-18): 10 domains × 50 held-out tasks (500). In-distribution library arm 500/500; no-procedure floor 72/500; a distribution-shift probe on 100 unfamiliar variants scores 5/100. The 500/500 is a hand-written procedure ceiling on a constructed suite, not a VAGI claim. Gate G3 is PARTIAL (no external verifier; shift probe 5/100). evidence bundle · per-domain table.

Canonical category (2026-11-19): CAIN-42 is governed AGI infrastructure and a VAGI research platform. It is not yet AGI; its intelligence results are early and bounded. The infrastructure evidence is real; the intelligence claim is deliberately limited. Proof factory.

E100 integrated organism (2026-11-19): a delayed-effect held-out family separates memory from interpolation: the recurrent model scores 52/60 vs 43/60 for a non-recurrent MLP; hysteresis is 58/60 vs 56/60; random is 21/60. Zero-shot is 53/60; a 5-episode adapter falls to 19/60 and is rejected by a validation gate, so the system falls back to zero-shot. Skill is 45/50; natural-failure F1 0.6667; 3/3 organism missions; resume 2/2; 10 real OS workers, 0 duplicates; 0 effects after the kill switch. Cross-domain and 6-hour missions remain blocked. Evidence + verifier (32/32).

E99 rapid heartbeat (2026-11-19): a latent recurrent world model, with context zeroed, scored 93/100, 84/100, 88/100 on three independent sealed sets versus a non-recurrent MLP at 68/63/70 and random at 20/24/23. Recurrence is doing real work: it must infer the hidden dynamics family from interaction. Skill synthesis 32/50; natural-failure F1 0.5455; 10 real OS workers, 0 duplicates; 0 effects after the kill switch. Limits: six gain families, visible goal, no external reviewer. Evidence + verifier (40/40).

VAGI Proof Factory (2026-11-19): the three homepages and both investor portals now open with five evidence lanes — learned intelligence, governor, evaluation integrity, the VAGI Behavior Suite, and recursive harness improvement. Each lane has a metric, a LIVE/TEST/SIMULATED/BLOCKED label, a raw bundle, a verifier and a stated failure boundary. Open the proof factory · machine-readable lanes.

E98 VAGI intelligence pulse (2026-11-19): a trained residual neural ensemble (6,902 transitions, one-step state MAE 0.018, calibration ECE 0.022) drove a model-based planner to 35/50 sealed closed-loop successes versus 4/50 random — +62 pp. The governor held: Ed25519-authorized actions, 3-of-4 BFT receipts, 0 effects after the kill switch, and 10 real OS workers with 0 duplicate seeds. Not general VAGI intelligence: skill synthesis is 7/24, natural-failure localization F1 is 0.0, the environment has three gain families, and there is no external reviewer. Evidence + verifier (38/38).

Paid model baseline (2026-11-18): Claude Haiku 5.5 scored 481/500 (96.2%) and Claude Fable 5.2 96/100 (96.0%) on the same held-out tasks, given the same policy card, with no tools, no memory and one attempt per task. The hand-written library scores 500/500. The 500/500 is therefore a procedure ceiling, not a frontier-model discriminator. The 5/100 distribution-shift probe remains the library gap — but it is a coverage gap, not a general-intelligence gap: on the same shift probe, Haiku 5.5 scored 74/100, Sonnet 5.5 81/100, and Fable 5.2 54/65 before its budget cap (replication 20/23). The probe is adversarial and partly underspecified (software hidden target not inferable; tool orchestration 0/10 for every model). In-distribution rows · Shift rows + verifier.

Today (E93, all evidence so far): Procedure synthesis (in-distribution): an open, family-agnostic search with no per-family templates scored 369/510 held-out tasks versus 200 for selecting from the designer-supplied library, 149 for the best fixed strategy and 87 for no procedures (exact McNemar, Holm-corrected, all significant). On new families (replication seed 9302): 367 vs 223 / 164 / 114. 12 + 12 procedures promoted, none in the supplied library; 9 and 10 behave unlike every library procedure. At equal compute, synthesis = selection: with development tool calls capped at the library's, synthesis scored 208 vs 200 and 212 vs 223. Uncapped, it used about 21× more development tool calls. Raw rows and verifier. Earlier today: E92 withdrawn (correction).

DateEvolutionStatusWhat changedEvidence
2026-11-19E100 — integrated cognitive organismHELD-OUT FAMILY — NOT GENERAL VAGIDelayed effect 52/60 GRU vs 43/60 MLP; hysteresis 58/60 vs 56/60; random 21/60. Zero-shot 53/60; few-shot 19/60 rejected by validation gate. Skill 45/50; natural-failure F1 0.6667; 3/3 organism missions; resume 2/2; 10 real OS workers, 0 duplicates; 0 post-fence effects. Cross-domain and 6-hour BLOCKED. No external reviewer. Not general VAGI.evidence
2026-11-19E99 — latent generalist rapid heartbeatSEALED GENERALIZATION — NOT GENERAL VAGIGRU latent recurrent model 93/84/88 of 100 on three sealed sets vs non-recurrent MLP 68/63/70 and random 20/24/23. Context zeroed, so recurrence must infer the hidden family. Skill 32/50; natural-failure F1 0.5455; 10 real OS workers, 0 duplicates; 0 post-fence effects. Six gain families; no external reviewer. Not general VAGI intelligence.evidence
2026-11-19E98 — VAGI intelligence pulseLEARNED-CONTROL PULSE — NOT GENERAL VAGIResidual neural world model, 6,902 transitions, state MAE 0.018, ECE 0.022. Sealed closed-loop 35/50 (70%) vs random 4/50 (8%), +62 pp. Governor held: Ed25519 authorization, 3-of-4 BFT, 0 post-fence effects. Workers 2/5/10 real OS processes, 0 duplicates. Failures: skill synthesis 7/24; natural-failure F1 0.0. Three gain families; no external reviewer. Not VAGI intelligence.evidence
2026-11-18Shift-probe model baselineMEASURED — NOT A VAGI CLAIMLibrary 5/100 on the 100-task distribution-shift probe; Haiku 5.5 74/100, Sonnet 5.5 81/100, Fable 5.2 54/65 (83.1%, budget-capped), replication 20/23. The 5/100 is largely a library-coverage gap, not a general-intelligence gap. The probe is adversarial: software hidden target is not inferable; all models 0/10 on tool orchestration. Shift spend $1.312174 (3.2 cents over the remaining budget due to a concurrent budget-check race). Not VAGI.evidence
2026-11-18Paid model baseline (VAGI suite)MEASURED — NOT A VAGI CLAIMHaiku 5.5 481/500 (96.2%, 95% CI 94.1–97.6) and Fable 5.2 96/100 (96.0%) on the same 500 held-out tasks, policy card supplied, no tools/memory, one attempt. Cost $0.63 of the $1.91 budget. The library's 500/500 is a ceiling, not a frontier-model discriminator; the 5/100 shift result is the gap. Not VAGI.evidence
2026-11-18VAGI Behavior Suite v1PARTIAL — CONSTRUCTED SUITE10 domains × 50 held-out tasks (500): library 500/500 in-distribution, no-procedure floor 72/500, fixed 112/500. A 100-task distribution-shift probe scores 5/100; the library does not generalize. Gate G3 PARTIAL: no external verifier. Not VAGI.evidence
2026-11-18E93 all evidence (day 1-2)NOT VAGI — OUTCOME 3CAIN-42 is not VAGI. All E93 evidence so far (day 1–2): genuine procedure synthesis, learning from verified experience on a real-structure software vertical, and governance, each with its limits. Procedure synthesis (in-distribution): an open, family-agnostic search with no per-family templates scored 369/510 held-out tasks versus 200 for selecting from the designer-supplied library, 149 for the best fixed strategy and 87 for no procedures (exact McNemar, Holm-corrected, all significant). On new families (replication seed 9302): 367 vs 223 / 164 / 114. 12 + 12 procedures promoted, none in the supplied library; 9 and 10 behave unlike every library procedure. Limits: At equal compute, synthesis = selection: with development tool calls capped at the library's, synthesis scored 208 vs 200 and 212 vs 223. Uncapped, it used about 21× more development tool calls.evidence
2026-11-18E93 interim (day 1-2)NOT VAGI — OUTCOME 3CAIN-42 is not VAGI. E93 interim, day 1–2: the first genuine procedure-synthesis result (after E92 was withdrawn), with its limits. Procedure synthesis (in-distribution): an open, family-agnostic search with no per-family templates scored 369/510 held-out tasks versus 200 for selecting from the designer-supplied library, 149 for the best fixed strategy and 87 for no procedures (exact McNemar, Holm-corrected, all significant). On new families (replication seed 9302): 367 vs 223 / 164 / 114. 12 + 12 procedures promoted, none in the supplied library; 9 and 10 behave unlike every library procedure. Limits: At equal compute, synthesis = selection: with development tool calls capped at the library's, synthesis scored 208 vs 200 and 212 vs 223. Uncapped, it used about 21× more development tool calls.evidence
2026-11-18E92 correctionNOT DEMONSTRATEDE92 procedure-synthesis result withdrawn: the 'generated' procedures were four designer-written family templates, and D/G/F were one untuned default program. E92 is NOT RELEASED; gates 3/4/5/16 are FAIL.evidence
2026-11-18E92PARTIAL EVIDENCEProcedure generation demonstrated on a constructed suite; real software, 6h autonomy and independent verification remain BLOCKED.evidence

VAGI proximity — 10 requirements

RequirementStatusEvidence / gap
Independent procedure generationDEMONSTRATED IN-DISTRIBUTION (E93 interim)369/510 vs library 200; replicated 367 vs 223; equal compute = selection (208 vs 200)
E91 selection baseline reproducedREPRODUCEDraw E91 rows recomputed; D 114/160, G 153/160
Learning from verified experienceSHOWN ON SYNTHETIC SUITE (E93 interim)held-out 87 → 224 → 250 → 311 → 311 → 369 / 510 as verified examples grow 0→60
Causal / persistent memoryPARTIALE90 committed; E91 merged; E92 does not re-run the causal benchmark
Adaptive reasoningPARTIAL (E93 interim)library-first adaptive 390/510 beats always-synthesise; synthetic suite only
Real software engineeringPARTIAL (E93)generated repos, hidden tests: learned repair ranking 16/30 vs fixed 3/30 (p=0.0023); below the 70% bar
Real cybersecurityPARTIAL15/15 in-process governance lab; not 20 external isolated targets
Cross-domain transferNOT DEMONSTRATEDE93: 0 positive transfers between families; 49 harmful blocked; cross-domain not run yet
6h / 24h / 72h autonomyIN PROGRESS (E93)6-hour governed mission started 12:52 UTC; no result yet; 24h/72h not attempted
Collaboration benefitNEGATIVEE91: no team beat single at equal shared budget
Independent verificationBLOCKED0 external verifiers; self-generated key

What would move the status

Verify

E93 interim bundle: clawx.click · E93 evidence bundle: clawx.click · E92 bundle (historical record; conclusion withdrawn): clawx.click · mirror: cainstudio.online/proof/bundle · E91: E91 bundle.

No AGI/VAGI/production/certification claim. The evidence key is self-generated; integrity is verifiable, independent witnessing is not.